Efficient Removal of Random Delays from Embedded Software Implementations Using Hidden Markov Models

Inserting random delays in cryptographic implementations is often used as a countermeasure against side-channel attacks. Most previous works on the topic focus on improving the statistical distribution of these delays. For example, efficient random delay generation algorithms have been proposed at CHES 2009/2010. These solutions increase security against attacks that solve the lack of synchronization between different leakage traces by integrating them. In this paper, we demonstrate that integration may not be the best tool to evaluate random delay insertions. For this purpose, we first describe different attacks exploiting pattern-recognition techniques and Hidden Markov Models. Using these tools and as a case study, we perform successful key recoveries against an implementation of the CHES 2009/2010 proposal in an Atmel microcontroller, with the same data complexity as against an unprotected implementation of the AES Rijndael. In other words, we completely cancel the countermeasure in this case. Next, we show that our cryptanalysis tools are remarkably robust to attack improved variants of the countermeasure, e.g. with additional noise or irregular dummy operations. We also exhibit that the attacks remain applicable in a non-profiled adversarial scenario. These results suggest that the use of random delays may not be effective for protecting small embedded devices against side-channel leakage. They highlight the strength of Viterbi decoding against such time-randomization countermeasures, in particular when combined with a precise description of the target implementations, using large lattices.